FLUX.2 Klein LoRA fine-tuning and image generation

This example demonstrates how to fine-tune a FLUX.2 Klein image generation model using LoRA (Low-Rank Adaptation) and generate images at scale on Deadline Cloud. The workflow trains a lightweight adapter from a small image dataset. It then uses that adapter to generate new images from text prompts, all orchestrated as Deadline Cloud jobs on a GPU fleet.

The source code for this example is available in the deadline-cloud-samples repository on GitHub.

The following video demonstrates the FLUX.2 Klein LoRA workflow on Deadline Cloud.

Estimated time: 30–60 minutes (including 15–45 minutes for training).

Overview

The example consists of two job bundles:

1. lora_training – Fine-tunes FLUX.2 Klein on your image dataset using LoRA, producing a .safetensors adapter file.

2. image_generation – Loads the base model with your trained adapter and generates images from text prompts. When generating multiple images, work is parallelized across workers.

LoRA is a parameter-efficient fine-tuning technique that trains a small adapter instead of modifying the full model weights. This approach makes training fast (15–45 minutes on a single GPU) and the resulting adapter portable and easy to version.

The job bundles use the diffusers and peft libraries for training and inference.

To complete this example, follow these steps:

1. Complete the prerequisites.

2. Prepare training data.

3. Submit the training job.

4. Submit the image generation job.

5. Clean up resources.

Prerequisites

Before you begin, make sure that you have the following:

• A Deadline Cloud farm with a GPU fleet (Linux, NVIDIA A10G GPU, 13 GB+ VRAM, 64 GiB+ system memory recommended).

• A queue associated with the GPU fleet.

• The Deadline Cloud CLI installed on your workstation.

• 20–50 training images of your subject (.jpg, .png, .jpeg, or .webp).

Job bundle structure

flux2_klein_lora/
├── lora_training/
│   ├── template.yaml       # OpenJD job template
│   └── train_lora.py       # Training script
└── image_generation/
    ├── template.yaml       # OpenJD job template
    └── generate_image.py   # Inference script

Prepare training data

To prepare your training data

1. Create a directory containing images of your subject. For best results, use 20–50 images with varied poses, lighting, and backgrounds.

2. Optionally, provide a .txt caption file alongside each image for more nuanced training:

   training_images/
   ├── IMG_001.jpeg
   ├── IMG_001.txt    # "a photo of ohwx dog sitting on grass"
   ├── IMG_002.jpeg
   ├── IMG_002.txt    # "a photo of ohwx dog running on beach"

   If no caption files are present, the training script generates them automatically using the instance prompt parameter. Custom captions produce higher-quality results because they teach the model more specific associations between your trigger word and visual features.

3. Choose a unique trigger word (such as ohwx). This binds the learned concept to a token that doesn't conflict with existing model knowledge.

Submit the training job

To submit the training job

1. Navigate to the flux2_klein_lora directory in your cloned deadline-cloud-samples repository.

2. To submit the training job with the GUI, run the following command:

   deadline bundle gui-submit ./lora_training

   Or, to submit the training job directly with the CLI, run the following command:

   deadline bundle submit ./lora_training \
     --queue-id gpu-queue-id \
     --parameter DatasetPath=~/training_images \
     --parameter InstancePrompt="a photo of ohwx dog" \
     --parameter OutputDir=/tmp/lora_output \
     --parameter MaxTrainSteps=1500 \
     --parameter Resolution=512

3. Monitor the job status in the Deadline Cloud console or by using the deadline job get command.

Training parameters

Training parameters

Parameter	Description	Default
Model Version	flux.2-klein-base-4b or flux.2-klein-4b	flux.2-klein-base-4b
Dataset Path	Directory containing training images	—
Instance Prompt	Text describing your subject	—
Resolution	Training image resolution in pixels	512
Network Dim	LoRA rank (higher = more capacity)	16
Network Alpha	LoRA alpha scaling factor	16
Max Training Steps	Number of training iterations	1500
Output Directory	Where to save trained weights	—

Training output

The training job produces the following artifacts:

• flux2_klein_lora.safetensors – The trained LoRA adapter with embedded metadata.

• Checkpoints saved every 300 steps.

To download the output after training completes, run the following command:

deadline job download-output \
  --job-id training-job-id \
  --queue-id gpu-queue-id

Submit the image generation job

To submit the image generation job

1. Navigate to the flux2_klein_lora directory in your cloned deadline-cloud-samples repository.

2. To submit the image generation job with the GUI, run the following command:

   deadline bundle gui-submit ./image_generation

   Or, to submit the image generation job directly with the CLI, run the following command:

   deadline bundle submit ./image_generation \
     --queue-id gpu-queue-id \
     --parameter LoRAPath=/tmp/lora_output/flux2_klein_lora.safetensors \
     --parameter Prompt="a photo of ohwx dog wearing a tuxedo" \
     --parameter OutputDir=/tmp/generated_images \
     --parameter NumImages=4

3. Monitor the job status in the Deadline Cloud console or by using the deadline job get command.

When NumImages is greater than 1, Deadline Cloud distributes the work across multiple workers for faster generation.

Generation parameters

Generation parameters

Parameter	Description	Default
LoRA Path	Path to trained .safetensors file	—
Prompt	Text description (include your trigger word)	—
Number of Images	Total images to generate	1
Width	Output width in pixels	1024
Height	Output height in pixels	1024
Inference Steps	Number of denoising steps	50
Guidance Scale	Classifier-free guidance scale	4.0

Generation output

Images are saved as image_0001.png, image_0002.png, and so on. To download them, run the following command:

deadline job download-output \
  --job-id generation-job-id \
  --queue-id gpu-queue-id

Model variants

FLUX.2 Klein model variants

Model	Parameters	Best for
flux.2-klein-base-4b	4B	Fine-tuning, commercial use
flux.2-klein-4b	4B	Fast inference (4 denoising steps)

Training recommendations

The following recommendations can help you get the best results:

• Dataset size – 20–50 high-quality images work well for most subjects.

• Resolution – Use 512 for GPUs with 24 GB VRAM. Higher resolutions require more memory.

• Training steps – Start with 1500. Increase if outputs don't resemble your subject.

• Network dim – 16 is a good default. Use 32 for complex or abstract concepts.

• Learning rate – The default of 1e-4 works well for most cases.

Clean up

To avoid ongoing charges, clean up the resources that you created for this example:

To clean up example resources

1. In the Deadline Cloud console, stop or delete the GPU fleet that you used for this example.

2. Delete any Amazon Simple Storage Service (Amazon S3) objects created by job attachments if they are no longer needed.

3. Remove local output files if they are no longer needed:

   rm -rf /tmp/lora_output /tmp/generated_images

Related resources

The following resources provide additional information:

• Example source code on GitHub

• FLUX.2 Klein on Hugging Face

• Open Job Description specification

• Diffusers documentation

• Deadline Cloud Architecture Guidance